This paper aims to develop a 3D scene graph representation that identifies the location and usage of functional interaction elements, enabling robots to directly interact with their environments. Instead of relying on traditional object-level resolution, we focus on detecting and storing objects at a finer resolution, focusing on their functional significance. To address data scarcity and the challenges of capturing detailed object features using robotic sensors, we leverage existing 3D resources to generate 2D data and train detectors, enhancing the standard 3D scene graph generation pipeline. Experimental results demonstrate that our approach achieves functional element segmentation performance comparable to state-of-the-art 3D models and enables task-driven functional semantic associations with higher accuracy than existing solutions.
